Nano Banana 2.1: Google's October Release AI Trackers Missed

What happened

One aggregator's tally of October 2026 model launches is incomplete. LLM Gateway's running list says one new AI model has arrived this month: luminal/DeepSeek-V4.1-Flash, dated October 4, 2026. Its seven-day view adds OpenAI's GPT-6.1 Sol from September 29. 4 Google's own changelog shows otherwise. On October 6, the Gemini API made Gemini Nano Banana 2.1 (model string gemini-nano-banana-2.1) generally available and deprecated the image model it replaces. 23

The omission matters because Nano Banana 2.1 is a production-grade GA release, not a preview or a minor patch. Releasebot's record of the Gemini API update describes it as a high-efficiency model for image generation and conversational editing. It is an update to Nano Banana 2, which shipped as gemini-3.1-flash-image. Google says it keeps Flash-tier speed and cost while improving visual quality, prompt adherence, character consistency across multi-turn edits, and text rendering. 3 It also adds wide and panoramic aspect ratios of 1:4, 4:1, 1:8 and 8:1 at multiple resolutions, starting at 1K and 2K. 3

Google's busy first week of October

The image model was only part of Google's activity. Google's release notes show the predecessor, gemini-3.1-flash-image, deprecated on October 6 with no shutdown date announced, and developers directed to migrate to Nano Banana 2.1. 2 Two days later, on October 8, Google deprecated two text-tier Flash models and set up automatic redirects. Requests to gemini-3.7-flash now route to gemini-3.8-flash, and requests to gemini-3.5-flash route to gemini-3.6-flash. Google suggests developers on the older model consider moving straight to 3.8 Flash. 2

The two sources agree on the core facts. Google's notes and Releasebot both give October 6 for the Nano Banana 2.1 launch and the 3.1 Flash Image deprecation. 23 Releasebot carries the fuller product description, and Google's page covers the follow-on Flash migrations. Together they show Google turning over several parts of its Gemini lineup within a few days.

Why the tracker missed it

The gap probably reflects how aggregators decide what counts as a "release" more than any single error. LLM Gateway lists what becomes available through its own API, and its entries are tied to the date a model was added to the platform. 4 An image-generation model, or a silent version bump hidden behind automatic routing, may never appear in a catalog built around text-model endpoints.

The same issue shows up in the one model the tracker does count. LLM Gateway presents DeepSeek V4.1 Flash as an October 4 release. 4 Tech Insider reports that DeepSeek actually published V4.1 Flash on HuggingFace on September 10, 2026, under an MIT license. 1 The October date fits a third-party provider, luminal, bringing the model onto the gateway, not DeepSeek's original launch. So the tracker undercounts Google's October and also moves a September open-weight release into October.

The open-weight side of the story

The DeepSeek release raises its own problems. According to Tech Insider, which credits reporting by Pasquale Pillitteri, independent developers uploaded "abliterated" and "uncensored" forks of V4.1 Flash with safety behavior removed within hours of the official upload. 1 By September 13, at least three such community builds were live, and the most popular had 2,254 downloads. 1 The outlet argues that this has become a repeatable pattern rather than a hypothetical risk. It adds that a month later, basic questions about who uploads these forks and whether anyone steps in remain unanswered. 1

These derivatives also show the limits of release tracking. A permissive license lets one upstream model spread into hosted endpoints, rebranded provider listings and modified forks, each with its own date. A tracker that records only one of these will misstate when, and in what form, the model actually became available.

The takeaway

In my reading, the "one model in October" figure tells you about one aggregator's catalog, not the market. Google shipped a GA image model and changed its Flash text tiers in the same week. DeepSeek's headline model is a September open-weight release whose descendants are still spreading. Developers who depend on model strings should treat vendor changelogs as the authoritative record. This applies especially when deprecations come with automatic rerouting, because traffic can quietly shift to a newer model without a code change. Third-party trackers work better as discovery tools than as a full ledger of releases.